def __call__(self, x):
if self.original_stride != 1:
_, _, width, height = cgt.infer_shape(x)
unstrided_width = width * self.original_stride[0]
unstrided_height = height * self.original_stride[1]
# workaround for this
# cgt.inc_subtensor(upsampled, (slice(None), slice(None), slice(None, None, self.original_stride[0])), slice(None, None, self.original_stride[1])), x)
placeholder = cgt.zeros((x.shape[0], x.shape[1], width,
unstrided_height)) # (None, 64, 4, 8)
cgt.inc_subtensor(placeholder,
(slice(None), slice(None), slice(None),
slice(None, None, self.original_stride[1])), x)
upsampled = cgt.zeros((x.shape[0], x.shape[1], unstrided_width,
unstrided_height)) # (None, 64, 8, 8)
cgt.inc_subtensor(
upsampled,
(slice(None), slice(None),
slice(None, None, self.original_stride[0]), slice(None)),
placeholder)
else:
upsampled = x
# then we conv to deconv
deconv = super(SpatialDeconvolution, self).__call__(upsampled)
# lastly we cut off original padding
pad = self.original_pad
original_width = (
(width - 1) * self.original_stride[0]
) - 2 * self.original_pad[0] + self.original_kernelshape[0]
original_height = (
(height - 1) * self.original_stride[1]
) - 2 * self.original_pad[1] + self.original_kernelshape[1]
t = deconv[:, :, pad[0]:(pad[0] + original_width),
pad[1]:(pad[1] + original_height)]
return t
def inc_row(M, x, Y):
if isinstance(M, np.ndarray):
M1 = M.copy()
M1[0:1] += x
else:
M1 = cgt.inc_subtensor(M, slice(0, 1), x)
return (M1 * Y).sum()
def inc_vec(M, x, Y):
if isinstance(M, np.ndarray):
M1 = M.copy()
M1[0] += x
else:
M1 = cgt.inc_subtensor(M, 0, x)
return (M1 * Y).sum()
def to_one_hot(y, nb_class, dtype=None):
"""
Return a matrix where each row corresponds to the one hot
encoding of each element in y.
Parameters
----------
y
A vector of integer value between 0 and nb_class - 1.
nb_class : int
The number of classes in y.
dtype : data-type
The dtype of the returned matrix. Default floatX.
Returns
-------
object
A matrix of shape (y.shape[0], nb_class), where each row ``i`` is
the one hot encoding of the corresponding ``y[i]`` value.
"""
fill_vals = cgt.ones((y.shape[0],))
ret = cgt.zeros((y.shape[0], nb_class), dtype)
d1 = cgt.arange(y.shape[0])
d2 = cgt.cast(y, 'i1')
ret = cgt.inc_subtensor(ret, [d1, d2], fill_vals)
return ret
def inc_row(M, x, Y):
if isinstance(M, np.ndarray):
M1 = M.copy()
M1[0:1] += x
else:
M1 = cgt.inc_subtensor(M, slice(0, 1), x)
return (M1 * Y).sum()
def inc_vec(M, x, Y):
if isinstance(M, np.ndarray):
M1 = M.copy()
M1[0] += x
else:
M1 = cgt.inc_subtensor(M, 0, x)
return (M1 * Y).sum()
def to_one_hot(y, nb_class, dtype=None):
"""
Return a matrix where each row corresponds to the one hot
encoding of each element in y.
Parameters
----------
y
A vector of integer value between 0 and nb_class - 1.
nb_class : int
The number of classes in y.
dtype : data-type
The dtype of the returned matrix. Default floatX.
Returns
-------
object
A matrix of shape (y.shape[0], nb_class), where each row ``i`` is
the one hot encoding of the corresponding ``y[i]`` value.
"""
fill_vals = cgt.ones((y.shape[0], ))
ret = cgt.zeros((y.shape[0], nb_class), dtype)
d1 = cgt.arange(y.shape[0])
d2 = cgt.cast(y, 'i1')
ret = cgt.inc_subtensor(ret, [d1, d2], fill_vals)
return ret
def test_incsubtensor0():
# First let's test fancy slice along zeroth dimension
W = cgt.shared(np.zeros((5, 3)), name="W")
inc = cgt.matrix() # we'll increment W by this matrix
incval = np.arange(9).reshape(3, 3)
inds = cgt.vector(dtype='i8')
updates = {W: cgt.inc_subtensor(W, inds, inc)}
f = cgt.function([inds, inc], [], updates=updates)
f([1, 2, 4], incval)
assert np.allclose(
W.op.get_value(),
np.array([[0., 0., 0.], [0., 1., 2.], [3., 4., 5.], [0., 0., 0.],
[6., 7., 8.]]))
def test_incsubtensor2():
W = cgt.shared(np.zeros((5,3)), name="W")
i0 = cgt.vector(dtype='i8')
i1 = cgt.vector(dtype='i8')
inc = cgt.vector()
updates2 = {W : cgt.inc_subtensor(W, (i0,i1), inc)}
f2 = cgt.function([i0,i1,inc],[],updates=updates2)
f2([0,1,2,2],[0,1,2,2],[1,2,3,4])
assert np.allclose(W.op.get_value(),
np.array(
[
[ 1., 0., 0.],
[ 0., 2., 0.],
[ 0., 0., 7.],
[ 0., 0., 0.],
[ 0., 0., 0.],
]))
def test_incsubtensor2():
W = cgt.shared(np.zeros((5, 3)), name="W")
i0 = cgt.vector(dtype='i8')
i1 = cgt.vector(dtype='i8')
inc = cgt.vector()
updates2 = {W: cgt.inc_subtensor(W, (i0, i1), inc)}
f2 = cgt.function([i0, i1, inc], [], updates=updates2)
f2([0, 1, 2, 2], [0, 1, 2, 2], [1, 2, 3, 4])
assert np.allclose(
W.op.get_value(),
np.array([
[1., 0., 0.],
[0., 2., 0.],
[0., 0., 7.],
[0., 0., 0.],
[0., 0., 0.],
]))
def test_incsubtensor1():
W = cgt.shared(np.zeros((5,3)), name="W")
inc = cgt.matrix() # we'll increment W by this matrix
incval = np.arange(9).reshape(3,3)
start = cgt.scalar(dtype='i8')
stop = cgt.scalar(dtype='i8')
updates = {W : cgt.inc_subtensor(W, slice(start, stop), inc)}
f = cgt.function([start,stop,inc],[],updates=updates)
f(0,3,incval)
assert np.allclose(W.op.get_value(),
np.array(
[
[ 0., 1., 2.],
[ 3., 4., 5.],
[ 6., 7., 8.],
[ 0., 0., 0.],
[ 0., 0., 0.],
]))
def test_incsubtensor1():
W = cgt.shared(np.zeros((5, 3)), name="W")
inc = cgt.matrix() # we'll increment W by this matrix
incval = np.arange(9).reshape(3, 3)
start = cgt.scalar(dtype='i8')
stop = cgt.scalar(dtype='i8')
updates = {W: cgt.inc_subtensor(W, slice(start, stop), inc)}
f = cgt.function([start, stop, inc], [], updates=updates)
f(0, 3, incval)
assert np.allclose(
W.op.get_value(),
np.array([
[0., 1., 2.],
[3., 4., 5.],
[6., 7., 8.],
[0., 0., 0.],
[0., 0., 0.],
]))
def test_incsubtensor0():
# First let's test fancy slice along zeroth dimension
W = cgt.shared(np.zeros((5,3)), name="W")
inc = cgt.matrix() # we'll increment W by this matrix
incval = np.arange(9).reshape(3,3)
inds = cgt.vector(dtype='i8')
updates = {W : cgt.inc_subtensor(W, inds, inc)}
f = cgt.function([inds,inc],[],updates=updates)
f([1,2,4],incval)
assert np.allclose(W.op.get_value(),
np.array(
[[ 0., 0., 0.],
[ 0., 1., 2.],
[ 3., 4., 5.],
[ 0., 0., 0.],
[ 6., 7., 8.]]))
